store.ts

1import { Database } from "bun:sqlite";2import { mkdirSync } from "node:fs";3import { join } from "node:path";4import { inflateSync } from "node:zlib";5import type { EventRecord, ReleaseRecord, SaveRecord } from "../bucket/repo.ts";6import { isTestPath } from "../describe/links.ts";78// Koh 3's schema (§5.1, `src/storage/storage.v`) minus `file_cache`, which only a working tree needs,9// plus `releases`: each version's release.json (§11.2) as read from the bucket, since that is bucket10// data too and pages and feeds need it on every request.11const SCHEMA_VERSION = 2;12const SCHEMA = `13CREATE TABLE objects (14	hash TEXT PRIMARY KEY,15	size INTEGER NOT NULL,16	data BLOB NOT NULL17);18CREATE TABLE saves (19	id            TEXT PRIMARY KEY,20	parent        TEXT REFERENCES saves(id),21	message       TEXT NOT NULL,22	created_at    INTEGER NOT NULL,23	machine       TEXT NOT NULL,24	is_checkpoint INTEGER NOT NULL DEFAULT 0,25	size_bytes    INTEGER NOT NULL DEFAULT 0,26	agent         TEXT,27	session       TEXT,28	restores      TEXT,29	replayed_from TEXT30);31CREATE TABLE save_files (32	save_id    TEXT NOT NULL REFERENCES saves(id),33	path       TEXT NOT NULL,34	hash       TEXT NOT NULL REFERENCES objects(hash),35	executable INTEGER NOT NULL DEFAULT 0,36	PRIMARY KEY (save_id, path)37);38CREATE TABLE events (39	id         TEXT PRIMARY KEY,40	kind       TEXT NOT NULL,41	save_id    TEXT,42	path       TEXT,43	reason     TEXT,44	created_at INTEGER NOT NULL,45	machine    TEXT NOT NULL46);47CREATE TABLE state (48	key   TEXT PRIMARY KEY,49	value TEXT NOT NULL50);51CREATE TABLE releases (52	version TEXT PRIMARY KEY,53	record  TEXT NOT NULL54);55CREATE INDEX saves_parent ON saves(parent);56CREATE INDEX save_files_hash ON save_files(hash);57`;5859/** An object as the bucket stores it: zlib bytes, already checked against its hash. */60export type StoredObject = { compressed: Uint8Array; size: number };6162/** One batch of a sync, landed in one transaction. Saves are oldest first. */63export type Landing = {64	head: string;65	releasesLatest: string;66	saves: SaveRecord[];67	objects: Map<string, StoredObject>;68	events: EventRecord[];69	releases: ReleaseRecord[];70	/** Escaped names of the paths these saves arrived without (§5.5, A3). */71	quarantined: string[];72	/**73	 * The operator's hidden paths (June, 2026-09-30), removed from every save like a strip, and the74	 * names ignored in every repo (June, 2026-10-01), each as `ANY_DEPTH` + name.75	 */76	hidden: string[];77};7879/** Marks a hidden entry as a name matched at any depth rather than a path from the root. */80export const ANY_DEPTH = "**/";8182// A recount inflates every file at the newest save; a file this large is data, not lines of code.83const MAX_COUNTED = 8 * 1024 * 1024;8485// Display names for the header's language, by extension; data and docs formats never count.86const LANGUAGES: Record<string, string> = {87	ts: "TypeScript",88	tsx: "TypeScript",89	js: "JavaScript",90	mjs: "JavaScript",91	zig: "Zig",92	v: "V",93	rs: "Rust",94	go: "Go",95	py: "Python",96	c: "C",97	h: "C",98	cpp: "C++",99	cc: "C++",100	hpp: "C++",101	swift: "Swift",102	kt: "Kotlin",103	java: "Java",104	rb: "Ruby",105	lua: "Lua",106	sh: "Shell",107	nix: "Nix",108	css: "CSS",109	html: "HTML",110};111112// Vendored code is someone else's: counting Koh 3's bundled SQLite made it read as 290k lines of C.113const VENDORED = /(^|\/)(vendor|third_party|node_modules)\//;114// Bumped when what counts as a line changes, so every mirror recounts once when it opens.115const STATS_RULE = "source-only";116117const open = new Map<string, Database>();118119/** The mirror of one repo at `<dir>/<repo>.db`, created on first use and kept open. */120export function openMirror(dir: string, repo: string): Database {121	const path = join(dir, `${repo}.db`);122	const cached = open.get(path);123	if (cached) return cached;124	mkdirSync(dir, { recursive: true });125	const db = new Database(path, { strict: true });126	db.run("PRAGMA journal_mode = WAL");127	db.run("PRAGMA foreign_keys = ON");128	db.run("PRAGMA busy_timeout = 5000");129	const { user_version } = db.query<{ user_version: number }, []>("PRAGMA user_version").get() ?? {130		user_version: 0,131	};132	if (user_version !== 0 && user_version !== SCHEMA_VERSION) {133		db.close();134		throw new Error(135			`${path} has mirror schema version ${user_version}; this Kepr reads version ${SCHEMA_VERSION}. ` +136				"The mirror is disposable: delete it and the next sync rebuilds it from the bucket.",137		);138	}139	if (user_version === 0) {140		db.transaction(() => {141			db.run(SCHEMA);142			db.run(`PRAGMA user_version = ${SCHEMA_VERSION}`);143		})();144	}145	// Mirrors counted under an older rule are recounted now, not at the next save.146	if (mirrorState(db, "head") && mirrorState(db, "stats_rule") !== STATS_RULE) refreshStats(db);147	open.set(path, db);148	return db;149}150151export function mirrorState(db: Database, key: string): string | null {152	const row = db153		.query<{ value: string }, [string]>("SELECT value FROM state WHERE key = ?")154		.get(key);155	return row?.value ?? null;156}157158export function knownSave(db: Database, id: string): boolean {159	return db.query("SELECT 1 FROM saves WHERE id = ?").get(id) !== null;160}161162export function knownObject(db: Database, hash: string): boolean {163	return db.query("SELECT 1 FROM objects WHERE hash = ?").get(hash) !== null;164}165166export function knownEventIds(db: Database): Set<string> {167	const rows = db.query<{ id: string }, []>("SELECT id FROM events").all();168	return new Set(rows.map((row) => row.id));169}170171/** Drops a version deleted from the bucket; a sync only ever adds releases. */172export function forgetRelease(db: Database, version: string): void {173	db.query("DELETE FROM releases WHERE version = ?").run(version);174}175176/** Replaces a version's record after Kepr rewrote it in the bucket. */177export function replaceRelease(db: Database, record: ReleaseRecord): void {178	db.query("UPDATE releases SET record = ? WHERE version = ?").run(179		JSON.stringify(record),180		record.version,181	);182}183184export function knownReleases(db: Database): Set<string> {185	const rows = db.query<{ version: string }, []>("SELECT version FROM releases").all();186	return new Set(rows.map((row) => row.version));187}188189/** Lands a sync like Koh's steal: objects, events, saves, then every strip over every save. */190/** Stores objects on their own, ahead of the save that refers to them. */191export function storeObjects(db: Database, objects: Map<string, StoredObject>): void {192	const put = db.query("INSERT OR IGNORE INTO objects (hash, size, data) VALUES (?, ?, ?)");193	db.transaction(() => {194		for (const [hash, object] of objects) put.run(hash, object.size, object.compressed);195	})();196}197198export function land(db: Database, landing: Landing): void {199	const putEvent = db.query(200		"INSERT OR IGNORE INTO events (id, kind, save_id, path, reason, created_at, machine) VALUES " +201			"(?, ?, ?, ?, ?, ?, ?)",202	);203	const putSave = db.query(204		`INSERT OR IGNORE INTO saves (id, parent, message, created_at, machine, is_checkpoint, size_bytes,205			agent, session, restores, replayed_from) VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)`,206	);207	const putFile = db.query(208		"INSERT OR IGNORE INTO save_files (save_id, path, hash, executable) VALUES (?, ?, ?, ?)",209	);210	const putRelease = db.query("INSERT OR IGNORE INTO releases (version, record) VALUES (?, ?)");211	const setState = db.query("INSERT OR REPLACE INTO state (key, value) VALUES (?, ?)");212	db.transaction(() => {213		storeObjects(db, landing.objects);214		for (const e of landing.events) {215			putEvent.run(e.id, e.kind, e.save_id, e.path, e.reason, e.created_at, e.machine);216		}217		for (const s of landing.saves) {218			putSave.run(219				s.id,220				s.parent === "" ? null : s.parent,221				s.message,222				s.created_at,223				s.machine,224				s.is_checkpoint ? 1 : 0,225				s.size_bytes,226				s.agent,227				s.session,228				s.restores,229				s.replayed_from,230			);231			for (const f of s.files) putFile.run(s.id, f.path, f.hash, f.executable ? 1 : 0);232		}233		for (const r of landing.releases) putRelease.run(r.version, JSON.stringify(r));234		applyStrips(db);235		let hidden = 0;236		for (const path of landing.hidden) hidden += removePath(db, path, "");237		if (hidden > 0) db.run("DELETE FROM objects WHERE hash NOT IN (SELECT hash FROM save_files)");238		// In the same transaction as the removal, so the tarball key never names a list the rows239		// don't reflect yet.240		setState.run("hidden", landing.hidden.join("\n"));241		setState.run("head", landing.head);242		setState.run("releases_latest", landing.releasesLatest);243		const quarantined = new Set([...quarantinedPaths(db), ...landing.quarantined]);244		setState.run("quarantined", JSON.stringify([...quarantined].sort()));245	})();246	// Outside the transaction: the recount can inflate megabytes and needs no lock to be correct.247	refreshStats(db);248}249250/** Escaped names of every path this mirror's saves arrived without. */251export function quarantinedPaths(db: Database): string[] {252	return JSON.parse(mirrorState(db, "quarantined") ?? "[]");253}254255// §10.2: a strip of one save reaches that save; a strip of every save bans its path from all of256// history, whenever a save was made or arrives. Re-applying every strip is harmless, which is how a257// save that arrives after its strip event is covered too.258function applyStrips(db: Database): void {259	const strips = db260		.query<{ save_id: string; path: string }, []>(261			"SELECT save_id, path FROM events WHERE kind = 'strip'",262		)263		.all();264	if (strips.length === 0) return;265	for (const strip of strips) removePath(db, strip.path, strip.save_id);266	db.run("DELETE FROM objects WHERE hash NOT IN (SELECT hash FROM save_files)");267}268269/** Removes a path, file or folder, from one save or ('') every save; returns the rows removed. */270function removePath(db: Database, path: string, save: string): number {271	if (path.startsWith(ANY_DEPTH)) {272		// A name ignored in every repo: any path with a segment of exactly that name.273		return db274			.query(275				`DELETE FROM save_files276				WHERE instr('/' || path || '/', $segment) > 0277				AND ($save = '' OR save_id = $save)`,278			)279			.run({ segment: `/${path.slice(ANY_DEPTH.length)}/`, save }).changes;280	}281	// Koh compares the prefix exactly, never with LIKE, so `_` and `%` in a path match only themselves.282	const dir = `${path.replace(/\/+$/, "")}/`;283	return db284		.query(285			`DELETE FROM save_files286			WHERE (path = $path OR substr(path, 1, length($dir)) = $dir)287			AND ($save = '' OR save_id = $save)`,288		)289		.run({ path, dir, save }).changes;290}291292/**293 * Empties a mirror in place so the next sync rebuilds it from the bucket, as a change to the294 * hidden paths needs. The file and its open handles stay, so a request holding one reads an empty295 * repo for a moment rather than failing. A new epoch keeps tarballs kept before it from matching.296 */297export function clearMirror(db: Database): void {298	db.transaction(() => {299		// Children before parents, for the foreign keys; every name is a table of SCHEMA's.300		for (const table of ["save_files", "events", "saves", "objects", "releases", "state"]) {301			db.run(`DELETE FROM ${table}`);302		}303		db.query("INSERT INTO state (key, value) VALUES ('epoch', ?)").run(crypto.randomUUID());304	})();305}306307// Lines and languages never change between syncs, so they are counted once here, not per view. They308// come from the newest real save, not `head`: a checkpoint's files are not public history (§21.4).309// Lines are source lines: files in a known language, tests left out; docs and data never count.310function refreshStats(db: Database): void {311	const save =312		db313			.query<{ id: string }, []>(314				"SELECT id FROM saves WHERE is_checkpoint = 0 ORDER BY rowid DESC LIMIT 1",315			)316			.get()?.id ?? "";317	const files = db318		.query<{ path: string; size: number; data: Uint8Array }, [string]>(319			"SELECT f.path, o.size, o.data FROM save_files f JOIN objects o ON o.hash = f.hash WHERE f.save_id = ?",320		)321		.all(save);322	let lines = 0;323	const byLanguage = new Map<string, number>();324	for (const file of files) {325		const language = LANGUAGES[file.path.split(".").pop() ?? ""];326		// Tests are not the program: counted, Koh 3's read as 18k lines of V where it has 11k.327		if (!language || VENDORED.test(file.path) || isTestPath(file.path)) continue;328		if (file.size > MAX_COUNTED) continue;329		const bytes = inflateSync(file.data);330		if (bytes.includes(0)) continue;331		let count = 0;332		for (const byte of bytes) if (byte === 10) count++;333		if (bytes.length > 0 && bytes[bytes.length - 1] !== 10) count++;334		lines += count;335		byLanguage.set(language, (byLanguage.get(language) ?? 0) + count);336	}337	const languages = [...byLanguage].sort((a, b) => b[1] - a[1]);338	const setState = db.query("INSERT OR REPLACE INTO state (key, value) VALUES (?, ?)");339	setState.run("lines", String(lines));340	setState.run("language", languages[0]?.[0] ?? "");341	setState.run("languages", JSON.stringify(languages));342	setState.run("stats_save", save);343	setState.run("stats_rule", STATS_RULE);344}